Data Science Engineer-I@ Techdome SolutionsHyderabad, India

I shipAI Solutions/Agentsto production.

Multi-agent systems, LLM fine-tuning and the reliability layer underneath. I build the agents — and the tooling that proves they work.

Career index
38
Public repos
2
PyPI packages
2.5+
Years shipping AI
5
Agents in prod system
LangGraphMulti-Agent OrchestrationRAGFine-TuningQuantizationOpenTelemetryPyTorchTransformersFAISSFastAPITGNNEvalsLangGraphMulti-Agent OrchestrationRAGFine-TuningQuantizationOpenTelemetryPyTorchTransformersFAISSFastAPITGNNEvals
01Experience

Four teams. One throughline: make AI work in production.

Selected roles, with the numbers that actually moved.

Data Science Engineer-I

Hyderabad
LangGraphLangChainFine-TuningQuantizationTGNN
  • 20% cheaper

    Cut coding-agent costs by fine-tuning and quantizing an open-source LLM on 100+ internal engineering docs, self-hosted on GPU servers (50%+ reduction in progress).

  • OSS

    Built and open-sourced AgentArgus, a framework-agnostic eval, observability & reliability layer — giving the team regression detection, tool-use scoring and failure-mode tracing before deployments.

  • 4–5 days → minutes

    Replaced a manual process with WizBot, a production agent generating 80–100 page business forms at 90%+ accuracy, auto-patching hundreds of live forms in seconds.

  • Zero direct DB

    Secured WizBot's data access with a guarded tool layer exposing only validated operations — eliminating direct DB access without losing functionality.

  • 5 agents

    Architected a LangGraph/LangChain system with orchestrator routing, checkpointing and memory, enabling reliable end-to-end execution of long-horizon agentic workflows.

  • Dual-learning

    Engineered a pipeline isolating client- and business-side updates, integrating TGNN for time-aware reasoning while preserving production model integrity.

02Flagship open source

Two libraries on PyPI. Both born from real production gaps.

Not tutorials — tooling I needed, packaged so other teams can use it.

P1Open source

AgentArgus

LLM-agent eval, observability & reliability framework

Python 3.10+OpenTelemetrySQLitePyPI
Problem

Teams ship agents blind — no visibility into whether they work, why they fail, or what they cost. Silent production failures, undetected regressions, and naive crash-or-retry on failed calls.

Impact

Wraps any agent and answers four questions: did it succeed, what did it cost, why did it fail, and can it recover. Framework-agnostic, no bundled LLM clients — you inject your own via a lightweight Judge protocol.

Evaluation

RAG metrics (faithfulness, answer relevance, context precision/recall) plus agent-behavior metrics — tool-use accuracy, plan coherence, success rate, error recovery.

Observability

OpenTelemetry tracing with GenAI semantic conventions, per-run token/cost accounting, trace-ID correlation and shareable HTML reports.

Reliability

Exponential-backoff retries, model fallback chains, circuit breakers and dead-letter queues — composed as breaker → fallback → retry.

Human-in-the-loop

Approval checkpoints that pause execution. Rejections are recorded failures, not crashes.

Orchestration

Supervisor/worker multi-agent pattern with durable SQLite checkpointing and resumable restarts.

example.pyready
from agentargus import Agent

# Wrap any callable — function, async fn, or BaseAgent
agent = Agent(my_agent, name="support-bot")

result = agent.run("Refund order #1024")

result.output      # the answer
result.cost        # per-run token accounting
result.trace_id    # OpenTelemetry correlation
result.tool_calls  # every tool invocation
result.scores      # RAG + behavior metrics

RunResult is the canonical immutable object — output, spans, costs, tool calls, steps, errors, scores.

P2Open source

methodoverload

True multi-dispatch method overloading for Python

PythonZero depsPyPI
Problem

Python lacks native method overloading, forcing brittle *args/type-check boilerplate. Hand-rolled dispatch logic repeats per project and makes multi-type constructors unreadable.

Impact

A single decorator dispatches one method name across multiple type signatures — restoring natural C++/Java-style overloading to idiomatic Python. Now used internally inside AgentArgus.

Type-safe dispatch

Resolves the correct implementation at runtime from standard type hints.

Full method support

Works on instance methods, classmethods and staticmethods via OverloadMeta.

Cached

Built-in signature caching keeps dispatch cheap on hot paths.

Clear failures

Explicit error messages when no overload matches, instead of silent fallthrough.

example.pyready
from methodoverload import overload

@overload
def add(a: int, b: int):
    return a + b

@overload
def add(a: str, b: str):
    return f"{a} {b}"

add(1, 2)               # 3
add("Hello", "World")   # "Hello World"

One name, many signatures — dispatched on type hints, with zero external dependencies.

02.cThe vault

Everything else, straight from GitHub.

Pulled live from the API — filter by language.

Snapshot · 38 repos
03Stack

Agent-first, top to bottom.

Not a flat keyword list — a connected graph. Hover any node to trace what it actually works with.

Building graph…
AgentsMLRetrievalDataInfraFundamentals
04Signals

Receipts.

Shipped outcomes, measured. Filter by kind, or read it as a timeline.

GitHub snapshot
38
Public repos
14
Stars earned
6
Languages shipped
2
PyPI packages
4–5 days → min

Shipped WizBot into active daily use at Techdome, collapsing business-form development from a 4–5 day manual process to minutes at 90%+ accuracy.

Techdome
1.5 hrs → 3 min

Cut a critical production job by ~30× through architectural refactoring, parallelization and bottleneck elimination, unblocking downstream workflows.

Sustainability Economics.ai
20% cheaper

Reduced coding-agent running costs by fine-tuning and quantizing an open-source LLM on 100+ internal engineering docs, self-hosted on GPU servers. 50%+ in progress.

Techdome
5-agent system

Architected a LangGraph/LangChain system with orchestrator routing, checkpointing and memory for reliable long-horizon agentic workflows.

Techdome
Zero direct DB

Eliminated direct database access from WizBot by designing a guarded tool layer exposing only validated operations — with no loss of functionality.

Techdome
Dual-learning

Engineered a pipeline isolating client- and business-side updates, integrating TGNN for time-aware reasoning while preserving production model integrity.

Techdome
2 packages

Authored and published AgentArgus (LLM-agent eval/observability/reliability) and methodoverload (Python multi-dispatch) on PyPI.

Open source
Top 8%

Ranked 16,684 of 200,000+ globally in the All India Common Internship Test.

AICIT
2nd in college

Held 2nd position on GeeksforGeeks within college for 3 consecutive weeks.

GeeksforGeeks
Top 20 / 250

Ranked top 20 of 250 interns for outstanding technical delivery and initiative.

DevTown
Education

B.Tech, Computer Science with AI

Pranveer Singh Institute of Technology, Kanpur

2021 — 2025GPA 7.01
04.bTrajectory
  1. Jul 2026Shipped

    AgentArgus v0.1.0 on PyPI

    Published the framework-agnostic agent eval, observability and reliability layer.

  2. Jan 2026Shipped

    methodoverload on PyPI

    Shipped true multi-dispatch method overloading for Python. Now used internally inside AgentArgus.

  3. Dec 2025Now

    Data Science Engineer-I — Techdome

    Agent infrastructure: WizBot, the 5-agent LangGraph system, guarded tool layers and cost reduction via fine-tuning + quantization.

  4. Sep 2025

    AI/ML Engineer — Sustainability Economics.ai

    RAG-grounded LLMs, the 1.5hr → 3min refactor, and standardized prompt operations on Azure AI Foundry.

  5. Feb 2025

    AI/ML Developer — Olibr

    AI Mock Interview System: resume parsing, semantic candidate matching with FAISS/Chroma, Flask inference APIs.

  6. Mar 2024

    Jr. Data Scientist — DevTown

    End-to-end ML models and data pipelines. Ranked top 20 of 250 interns.

  7. 2021 — 2025

    B.Tech, CS with AI — PSIT Kanpur

    Computer Science with an AI specialization. GPA 7.01.